Skip to content

Releases

v0.29.0

Choose a tag to compare

@niksa90 niksa90 released this 19 Nov 16:34
Immutable release. Only release title and notes can be modified.
dfcd28d

What's Changed

  • Retire description/late-instruction by @niksa90 in #82
  • 0.29.0: instructions/unknown-name, spec/discover-refused, config rule-id warning, auth login and diff notes by @niksa90 in #83

Full Changelog: v0.28.1...v0.29.0

v0.28.1

Choose a tag to compare

@niksa90 niksa90 released this 17 Nov 21:12
Immutable release. Only release title and notes can be modified.
7f31609
Merge pull request #80 from niksa90/fix/auth-login-no-login-needed

auth login and connect: no login wall where the server asks for none

v0.28.0

Choose a tag to compare

@niksa90 niksa90 released this 11 Nov 13:25
Immutable release. Only release title and notes can be modified.
e278bae

What's Changed

  • snapshot --read-resources: read each listed resource by @niksa90 in #75
  • apps/resource: run failures are info; the weekly bench reads resources by @niksa90 in #76
  • 0.28.0: version, changelog, Action pins; the PR comment checks toolmenu mcp by @niksa90 in #77

Full Changelog: v0.27.1...v0.28.0

v0.27.1

Choose a tag to compare

@niksa90 niksa90 released this 10 Nov 07:42
Immutable release. Only release title and notes can be modified.
636155b

What's Changed

  • README as the pitch, the reference in docs/; two more MCP tools; the plugin on 0.27.0 by @niksa90 in #73
  • 0.27.1: version, changelog, Action pins by @niksa90 in #74

Full Changelog: v0.27.0...v0.27.1

v0.27.0

Choose a tag to compare

@niksa90 niksa90 released this 09 Nov 21:50
Immutable release. Only release title and notes can be modified.
8d2e3fb

What's Changed

  • 0.27.0: review, auth check, call every tool, more clients, and a weekly bench by @niksa90 in #72

Full Changelog: v0.26.0...v0.27.0

v0.26.0

Choose a tag to compare

@niksa90 niksa90 released this 08 Nov 22:11
Immutable release. Only release title and notes can be modified.
5da9c76

Security

From a security audit of the repository.

  • auth login opens only https login addresses (or http to this machine).
    The address comes from the server's OAuth metadata, and the browser opener does
    whatever its scheme says: a hostile server could have xdg-open, open or
    rundll32 open a file:, smb: or app link, which on Windows can start a program.
    It's refused before anything is opened or printed.
  • toolmenu mcp: honest annotations, save_as inside cwd. snapshot_menu and
    check_session run the command they're given, but were annotated
    destructiveHint: false, so a client that auto-approves non-destructive tools would
    let a prompt-injected agent run anything. They're destructiveHint: true now.
    save_as writes only inside cwd (symlinks followed): a new file, or an existing
    menu file, never a shell profile or a source file.
  • Reports show a server's text; they never act on it. Tool names, descriptions,
    serverInfo and error text went into reports as sent. In a terminal, escape
    sequences could change the title, clear the screen or fake links, and a newline
    could forge a finding or a ::error:: line in a CI log. In markdown (the Action's
    PR comment) a backtick ended the code span, and images, links, @mentions and
    #references rendered, posted by github-actions. Now control, invisible and
    text-direction characters are shown as escapes (\u001b, \n), markdown is escaped
    around toolmenu's own code, and every write to stdout and stderr goes through the
    same escaping. JSON output stays the same JSON. The menu file keeps the exact bytes.
  • The Action's PR comment can't be taken over. It edited the first comment with
    its marker, whoever wrote it; now only one its own token wrote.
  • The Action keeps PR text out of its arguments and its markdown: --release is
    one argument (a package.json version with spaces added arguments), and error
    output goes in a code block its backticks can't close.
  • scan fills ${VAR} only from variables you name with --allow-env (or '*').
    A config decides where each value goes, a URL included, so scanning someone else's
    .mcp.json could send any secret in your environment to their server.
  • Credentials off the command line. TOOLMENU_ARGS (a JSON array of options and
    their values) keeps --header, --env and --client-secret values out of ps,
    where every user of the machine can read them; the Action passes its headers and
    env that way (to an older version still as arguments, with a warning).
  • Release and CI workflows: least-privilege tokens, every action pinned to a
    commit (Dependabot keeps them current), and the npm package built in a job apart
    from the tests and published from a job that runs no dependency, with npm's trusted
    publishing: no long-lived npm token anymore. toolmenu init
    pins the actions it writes the same way.
  • Smaller: a menu file, toolmenu.config.json or scan list that isn't JSON is
    reported without quoting it; JUnit output escapes characters a terminal acts on;
    history uses only semver-shaped registry versions; macOS runners (bash 3.2) get
    their SARIF files.

Breaking

  • scan: a config that uses ${VAR} needs --allow-env VAR (or --allow-env '*'
    for the old behaviour). The entry says which.
  • toolmenu mcp: save_as outside cwd, or over a file that isn't a menu, is an
    error. The release-check skill keeps its baseline inside the worktree.
  • snapshot exits 1 on a server whose tool names hold control, invisible or
    text-direction characters (spec/tool-name, below).

Also

  • spec/tool-name: a tool name outside the format the spec recommends from
    2025-11-25 (1 to 128 of A-Z a-z 0-9 _ - .) is a warning (info on older
    protocols), and an error when it hides characters.
  • An HTML report. --html <path> writes the run of snapshot, diff or session
    as one self-contained page beside the usual report: stat cards (errors, warnings,
    tools, tokens; breaking changes and the suggested bump for diff), the menu as a
    table in server order with each tool's share of the tokens, and the findings with
    their next steps. diff shows each tool before and after; session has a row per
    step with the menu after it, marking tools inserted, appended or edited, and links
    each step to its findings. No script and nothing fetched, so it opens offline from a
    CI artifact; light and dark follow the viewer. The design follows the Board UI
    design system, in the logo's colours.
  • Library: session takes onStepMenu (the menu after each step, with what the step
    inserted, appended, moved, edited or removed; stepMenu and StepTool are
    exported), kept out of the result so the JSON output is unchanged. New
    comparedTools(before, after, ignore) returns the tools diffMenus compares (after
    the ignore globs, with their served positions) and which were added, removed and
    renamed; diffMenus uses it too.

v0.25.2

Choose a tag to compare

@niksa90 niksa90 released this 08 Nov 07:59
79f0639

What's Changed

  • JUnit output for CI test reports; document the library API and routes.yml by @niksa90 in #66
  • Reject HTTP server URLs with embedded credentials in serverArgs by @niksa90 in #67
  • 0.25.2 by @niksa90 in #68

Full Changelog: v0.25.1...v0.25.2

v0.25.1

Choose a tag to compare

@niksa90 niksa90 released this 06 Nov 01:24
45d002c
  • In the MCP Registry. server.json lists toolmenu's MCP server as
    io.github.niksa90/toolmenu, and package.json gains the matching mcpName, which
    is how the registry checks that the npm package is this server's. The release
    workflow publishes the entry after npm, signed in with GitHub's OIDC token.
  • The npm page shows the logo and the Glama badge

0.25.0

Choose a tag to compare

@niksa90 niksa90 released this 05 Nov 21:14
884db71
  • A README badge. snapshot --badge <path> writes a shields.io endpoint file
    (toolmenu | 12 tools · ~4.1K tokens · 2 warnings, red with errors, yellow with
    warnings) and prints the README line, which reads it from GitHub on the default
    branch. Committed with the baseline, it changes when the baseline does. toolmenu's
    README shows its own MCP server's.
  • SARIF output, for GitHub code scanning. --sarif <path> writes the findings
    of snapshot, diff or session as SARIF 2.1.0 beside the usual report;
    --format sarif prints it instead. Each result points at the menu file, at the line
    of the tool it's about (--sarif-menu names the committed one when the menu was
    written elsewhere), with a fingerprint from its rule and tool (and which of that
    rule's findings on the tool it is) so alerts track across runs. The Action's new sarif input writes snapshot.sarif, diff.sarif and
    session.sarif into a folder for github/codeql-action/upload-sarif.
  • Plugin: two more skills, and evals. release-check (before tagging or
    publishing: what changed since the last release, what breaks clients, the bump it
    needs) and setup-ci (adds toolmenu to CI with toolmenu init). check-mcp-menu
    names more of the moments it's for (a rename, a description edit, an agent picking
    the wrong tool) and falls back to the CLI without the MCP tools.
    plugins/toolmenu/evals/ has seven cases for claude plugin eval: each skill fires
    on its own prompts (3 of 3 runs, 0 of 3 without the plugin), none fires on unrelated
    work, and setup-ci gives the right init command for a scaffolded project.

v0.24.0

Choose a tag to compare

@niksa90 niksa90 released this 05 Nov 18:51
2ad9847

What's Changed

  • 0.24.0: toolmenu as an MCP server, and a Claude Code plugin by @niksa90 in #59

Full Changelog: v0.23.0...v0.24.0